Converting the output tensors into floats for the prost-processors
causes memory exhaustion when models become very large. Additionally,
it wastes processing time converting values which may not be used. By
moving the conversion step into the post-processors we avoid this issue.
If no callback is passed for post-processing the converted output to
a floating point ndarray is returned still.
Originally meant to abstract gc_collect but we could just use
m_alloc and friends. Also was meant to provide functions like
alloc0, alloc_maybe etc.. which are all available in MP anyway.
Signed-off-by: iabdalkader <i.abdalkader@gmail.com>